👋 Hi, I’m Andre and welcome to my newsletter Data Driven VC which is all about becoming a better investor with data and AI.
Upcoming events:
‘The Agentic Fund’ webinar on Oct 28 with Balderton & Harmonic
DDVC x Slush Breakfast in Helsinki Nov 18 with Goodwin & Harmonic
Brought to you by VESTBERRY - Portfolio Management Platform for data-driven VCs
As Giant Ventures launched new funds and its portfolio grew, keeping everything in sync took more and more time. Valuations were updated by hand in Excel, and the information behind them was scattered across files.
Vestberry brought it all into one place. Quarterly valuations and LP reports became much faster. Audits got simpler too, now that the team can pull the portfolio position for any past date.
Welcome to another Data Driven VC “Insights” episode where we cover the most interesting research and reports about startups, GPs, LPs, AI & automation.
Why VCs Sell Their Winners After IPO
Dan Gray at Odin tracks how fast VCs sell after IPO and argues that early-release clauses let them cash out while the market is still hot.
Only 8 of 50 Beat the Index: Of the 50 largest US VC-backed IPOs since 2014, only 8 have outperformed the Nasdaq-100 (QQQ) since listing, against 14 of 50 for 1994-2006 and 22 of 50 for 2004-2016.
VCs Start Selling Twice as Soon: Across 386 venture stakes, investors in 2014 IPOs made their first sale a median 14.7 months after listing. Since 2017, first sales have come at 5.9 to 6.9 months, around when the standard 180-day lockup (a post-IPO sales ban) ends.
Early Selling Became Standard: Early-release clauses let insiders sell before the lockup ends if the stock clears a set price at first earnings. In Gray's sample, they rose from 0% of VC-backed IPOs in 2016 to 83% in 2021 and still appear in about 80%.
✈️ KEY TAKEAWAYS
Gray's point is credibility: VCs who sell into hot markets while calling public markets short-sighted make IPO buyers warier, which he suggests may be why going public now feels harder. GPs who want a healthier IPO window may have to show conviction by holding after listing.
If Intelligence Is Getting Cheap, What Remains Expensive?
Prathik Naidu, founding partner at Long Lake, argues that buying services businesses outright is the path to putting AI to work.
Owning Outcomes Stays Expensive: As intelligence gets cheaper, Naidu argues the costly part is being accountable for results. Long Lake targets services firms that clients trust end to end, where the work needs human judgment and regulation sets who is responsible.
The Client Keeps the Upside: Outside vendors must clear security reviews, drive adoption and set pricing. Even then, a $1M-a-year contract in Naidu's example adds $5M of yearly EBITDA for the client, worth $50M at a 10x multiple, while the vendor must keep winning renewals.
Ownership Puts AI Into the P&L: Building an AI-native firm from scratch means starting without customers, brand or expertise, so Long Lake buys established businesses. Its AI rollouts cost first, then free up staff time, which Naidu says reaches full value only once it is put to work: higher revenue per employee and better service.
✈️ KEY TAKEAWAYS
For VCs, the piece raises a hard question: if AI vendors leave most of the gain in the client's P&L, should funds buy services businesses and run the AI rollout themselves? If this framework is right, the biggest AI returns in services will go to whoever owns the business.
Upgrade your subscription to access our premium content & join the Data Driven VC community
The Follow-On Trap
Ilya Strebulaev at Stanford GSB explains how escalation of commitment, the urge to keep investing in a failing project, distorts follow-on decisions inside VC partnerships.
The Closest Partner Overcommits: The lead partner usually sits on the board and knows the company best. Strebulaev argues that closeness, plus the wish to prove the first call right, makes them too upbeat to judge the next round.
Sunk Costs Drive Bad Bets: In his executive education test, a project has absorbed $500M with no profit. Often 30% or more choose to add $100M for a 5% chance at $1B, a bet worth far less than its cost.
Outside Leads Break the Cycle: Almost nine in ten firms use the same rules for follow-ons as for new deals. The piece suggests a “red team” that excludes the deal lead, and new investors leading rounds, which his research links to higher fund returns.
✈️ KEY TAKEAWAYS
GPs should look back at past follow-ons and limit the deal lead's say in the next one, since the partner closest to the company is the most prone to overcommit. Founders should treat an insider keen to lead their next round alone as a signal worth questioning, and keep courting outside investors who could lead it.
Agent Infra Is AI's Third Wave
Niklas Falk and Tobias Winczer at Balderton argue that long-running agents have created a third wave of AI infrastructure, following model training and chatbots.
Each Workload Needs New Infra: Model training needed GPU clusters and training data, and chatbots moved the load to inference and live web search. Agents are the third workload, running for hours, hundreds or thousands at once, and acting inside company systems.
Agents Need Their Own Stack: Earlier infrastructure was built for chatbots that answer one request and move on. Balderton maps agent infra into six layers: where agents run (including sandboxes for code), context, tool use, orchestration of long tasks, self-improvement, and security and observability.
Four Forces Decide Which Layers Pay: Strong adoption does not always bring revenue. The team names four forces that decide which layers become large businesses: (1) customers building parts in-house, (2) AI labs bundling what is on their roadmap, (3) data platforms selling automation on top of customer data and (4) clouds moving up the stack.
✈️ KEY TAKEAWAYS
Balderton sees value move from compute toward the layers closest to the work, so after inference the next winners sit in agent execution, such as sandboxes and keeping long tasks running. For investors, the best bets are layers that none of the four forces can easily absorb.
Join 2,023+ investors in our free Slack group as we automate our VC job end-to-end with AI. Live experiment. Full transparency.
How to Use Metrics in a Fundraise
Aaron Harris, founder of Magid and former YC partner, explains how to use metrics in a fundraise when investors share no common benchmarks.
Know Your Numbers Cold: Harris calls not knowing your metrics disqualifying. Founders should avoid flattering tricks such as cumulative charts or GMV reported as revenue, know how each number is calculated, and when stuck, promise an answer the same day.
Benchmark Against the Queue: With no shared benchmarks, investors judge you against the other companies in their pipeline, and AI's revenue speed has raised the bar. Harris advises timing the raise carefully and asking investors which companies they passed on or missed.
Show a Trend: Six months of 40% monthly growth reads as an early trend. Harris flags three weak spots: (1) too little data, (2) months that swing between growth and decline, which need an honest structural explanation, and (3) a long flat history before a recent spike.
✈️ KEY TAKEAWAYS
Founders should lead with the metric that removes the biggest doubt about the business, such as enterprise sales for a team that has never sold to enterprises. Investors can apply the same test in diligence: find the biggest risk, then check how the metric addressing it is calculated.
Long-Running Agents Need Durable State
Avi Chawla, co-founder of Daily Dose of Data Science, builds an AI agent that pauses for human approval and resumes hours later, using Google's agent tools.
Deploying Is Only the Start: Chawla argues a production agent must outlast the program that started it. An agent waiting hours for a human sign-off may be shut down, and simply rerunning the request can change its evidence and draft, or publish twice.
Save the Whole Unfinished Task: The fix is to store the paused task, the documents it found, its draft and the pending approval. In Chawla's test, a new program picked the task up later, published from the original draft and ran the search only once.
Approvals Must Not Run Twice: A resumed agent can repeat a step, which is harmless for a search but costly for a payment or email. Chawla tags each action with a unique key, so a repeated approval simply returns the first result.
✈️ KEY TAKEAWAYS
Anyone running agents that pause for approval should check two things: can a paused task survive a restart, and can a repeated approval trigger the same action twice?
That’s it for today!
Stay driven,
Andre
